run:
  name: discovery_benchmark
  output_dir: results/discovery_benchmark
  random_seed: 20260415
  batch_size: 32
  adaptive_budget: 50000
  baseline_budget: 50000
  max_batches: 2000
  allow_resume: true
  enable_adaptive_early_stop: true

target:
  protein_name: MDM2
  target_id: mdm2_discovery
  docking_reference_pdb: "4HG7"
  docking_target_path: data/targets/discovery_benchmark/mdm2_4hg7.pdb

reference:
  reference_id: ref_nutlin3a
  pdb_id: "4HG7"
  ligand_comp_id: "NUT"

benchmark_dataset:
  output_dir: data/ligands/discovery_benchmark
  target_size: 50000
  shuffle_seed: 424242
  reuse_existing: true
  min_similarity_keep: 0.08
  pubchem_max_records: 120000
  pubchem_thresholds: [95, 92, 90, 88, 85, 82, 80, 77, 75, 72, 70, 67, 65, 62, 60, 57, 55, 52, 50, 47, 45, 42, 40, 37, 35, 32, 30]
  chembl_target_id: CHEMBL5023
  chembl_max_rows: 120000
  allow_generated_fallback: true

backend:
  type: rdock
  require_real_backend: true
  n_runs: 5
  mapper_radius: 6.0
  command_timeout_seconds: 300
  parallel_jobs: auto-minus-4

encoding:
  fingerprint_radius: 2
  fingerprint_bits: 1024
  generate_3d: false

feature_extraction:
  compute_partial_charges: false
  compute_sasa: false

clustering:
  butina_cutoff: 0.35
  n_hyperclusters: 50

scheduler:
  init_coverage_fraction: 0.35
  conservative_deprioritize: true
  surrogate:
    prefer_xgboost: true
    n_estimators: 250
    min_train_samples: 20
    max_depth_small: 3
    max_depth_large: 6
  model_weight_schedule:
    sample_knots: [20, 50, 100, 200, 500]
    weight_knots: [0.1, 0.3, 0.5, 0.7, 0.9]
    max_weight: 0.9
    min_weight: 0.05
    instability_threshold: 2.0
    instability_decay: 0.25

matrix:
  naive_random_seeds: [11, 22, 33, 44, 55]
  cluster_naive_seeds: [101, 202, 303, 404, 505]
  include_adaptive_top1_variant: true
  include_adaptive: true
  include_naive_random: true
  include_cluster_naive: true
  modes: [docking_only, full_feature]

analysis:
  analog_similarity_threshold: 0.65
  rolling_window: 200
  replicate_subset_size_top: 20
  replicate_subset_size_borderline: 20
  replicate_subset_size_clusters: 20
  replicate_runs: 2

early_stop:
  min_evaluated: 3000
  patience_rounds: 12
  min_model_weight: 0.4
  quantile: 0.15
  uncertainty_weight: 0.2
  min_improvement: 0.015
  exploration_margin: 0.45
  exploitation_margin: 0.1
  reference_evaluations: 50000